Dimma: Semi-supervised Low Light Image Enhancement with Adaptive Dimming

arXiv

Dimma: Semi-supervised Low Light Image Enhancement with Adaptive Dimming

Wojciech Kozłowski, Michał Szachniewicz, Michał Stypułkowski, Maciej Zięba

Abstract

We propose Dimma, a semi-supervised approach that aligns with any camera by utilizing a small set of image pairs to replicate scenes captured under extreme lighting conditions taken by that specific camera. Dimma is the perfect solution for swiftly crafting a customized low-light image enhancement model for your camera, all without requiring an extensive collection of image pairs.

Results

Few-shot Dark

Method	PSNR ↑	SSIM ↑	RGB-SSIM ↑	LPIPS ↓	DeltaE ↓	NIQE ↓
SNR-Net	19.43	0.78	0.75	0.42	9.59	4.61
LLFlow	19.46	0.81	0.79	0.35	9.69	3.50
Dimma (ours)	24.14	0.83	0.81	0.27	6.14	2.93

LOL

Method	PSNR ↑	SSIM ↑	RGB-SSIM ↑	LPIPS ↓	DeltaE ↓	NIQE ↓	Train pairs ↓
KinD++	21.80	0.88	0.83	0.16	8.50	4.00	460
SNR-Net	24.61	0.90	0.84	0.15	6.85	4.02	485
LLFlow	25.19	0.93	0.86	0.11	6.40	4.08	485
Dimma 3 pairs	23.54	0.83	0.76	0.26	9.20	3.93	3
Dimma 5 pairs	24.49	0.84	0.76	0.25	7.98	3.81	5
Dimma 8 pairs	24.70	0.86	0.78	0.23	7.81	3.56	8
Dimma full	27.39	0.91	0.86	0.11	5.54	3.14	480

Run the code

To run this code install requirements

pip install -r requirements.txt

and run the following commands:

python train_unsupervised.py --config="configs/LOL/stage1/3shot-lol.yaml"
python finetune.py --config="configs/LOL/stage2/3shot-lol-ft.yaml"

For different config file use --config flag. There are many configs in config folder.

Please, bear in mind that you need to first train unsupervised model before running finetune.py. Data and models are not included in this repository. You can get them from the following link: drive.

Citation

If you find our work useful for your research, please cite our paper

@article{kozlowski2023dimma,
  title={Dimma: Semi-supervised Low Light Image Enhancement with Adaptive Dimming},
  author={Koz{\l}owski, Wojciech and Szachniewicz, Micha{\l} and Stypu{\l}kowski, Micha{\l} and Zi{\k{e}}ba, Maciej},
  journal={arXiv preprint arXiv:2310.09633},
  year={2023}
}

WojciechKoz/Dimma